Objective   A rapid and accurate detection of the fault occurring to the airborne electronic system plays a crucial role in ensuring the safety of civil aircraft. However, due to the increase of circuit board size and component density in airborne electronic system, the traditional contact fault diagnosis method encounters various problems such as low accuracy, huge time cost and the demanding requirements on personnel competency. Therefore, this study aims to explore the solution to circuit board fault diagnosis based on non-contact infrared technology, which is essential for improving the accuracy of fault diagnosis for the airborne electronic system.

  Methods   After the sequential thermal image of the circuit board is captured by using the infrared camera, the region of interest in the thermal image is processed as the infrared temperature series. Since the infrared temperature series of the circuit board contains various fault-related information, the accuracy of fault diagnosis can be improved by making full use of its local and global features. In this study, a fault diagnosis algorithm is proposed to achieve this purpose. Composed of the features extraction network (FEN) and the relationship learning network (RLN), it utilizes the local features of temperature series and the relationship between the features. Built on a residual structure with multi-scale dilated CNN, FEN plays the role of a local-feature extraction network to construct a multi-scale receptive field without increasing the number of training parameters and to learn the spatial features of temperature series of different ranges. Based on the embedded structure of two identical layers, attention mechanism and LSTM network, RLN is a network that can apply control on the transmission of temperature series to learn the importance of features and assign attention weights for mining the correlations between the features extracted from different positions. To develop a complete circuit board fault diagnosis algorithm, the parallel FEN and RLN networks are connected to the "Softmax" classifier.

  Results and Discussions   The temperature series datasets representing 27 different fault categories are constructed based on the infrared thermal image of airborne power board (Tab.1, Tab.5). (1) By analyzing the temperature series datasets, it can be found that there are significant differences between the temperature curves of the chip under different fault conditions, and the temperature curves of non-faulty chips are also affected by faulty chips (Fig.5). (2) The experimental results show that the proposed algorithm achieves a better diagnostic performance than FCN, MFCN, LSTM and LSTM-FCN on the datasets of the temperature series testing on two self-built circuit boards. To be specific, its diagnostic accuracy reaches 91.15% and 96.27%, respectively (Fig.8) (Tab.5). (3) Given the identical hyperparameter setting, the increase in dimension of temperature sequence feature vector contributes to improving the diagnostic performance. That is to say, appropriate sample is one of the key influencing factors in improving the accuracy of fault diagnosis (Tab.5). (4) Ablation studies reveal that the performance of FEN in feature extraction capability can be improved by the proper setting of hyperparameters, which is conducive to enhancing the diagnostic accuracy of the algorithm (Tab.6). (5) The long Short-term Memory hybridized with Attention (LSTMwAtt) plays a role in improving the performance of the proposed algorithm in terms of relation extraction. By fully utilizing the intrinsic relationship between the characteristics of different locations of temperature series, the proposed algorithm is more likely to capture the differentiated data carried by similar faults (Tab.6).

  Conclusions   In this study, a fault diagnosis algorithm intended for the airborne circuit board is proposed by using infrared temperature series. In this algorithm, the features extraction network is responsible for extracting local features and learning the spatial features of temperature series of different ranges, while the relationship learning network is proposed to discover the intrinsic relationships among the representations learned from infrared temperature series. According to the experimental results, the proposed diagnosis algorithm performs well on self-built testing datasets. However, it is worth noting that the small size of the self-built datasets reduces the accuracy of the algorithm when the proposed algorithm is applied to the new datasets. As the size of self-built datasets increases, it performs better in fault diagnosis. Hopefully, it would be applicable in circuit board fault systems to deal with the fault that occurs to the airborne electronic systems.